🕵️♂️(he/him)
Marie Skłodowska-Curie Postdoctoral Fellow
I am a Marie Skłodowska-Curie (MSCA) Postdoctoral Fellow at the Artificial Intelligence Research Institute (IIIA), Spanish National Research Council (CSIC), collaborating with Felip Manyà on my MSCA research project Sherlock4Py. Previously, I was a MSCA COFUND Postdoctoral Fellow at IIIA–CSIC, collaborating with Felip Manyà and Guillem Alenyà (IRI-CSIC-UPC) on the ALLIES project.
Before joining IIIA-CSIC, I was a Postdoctoral Research Associate at the University of Oxford, contributing to the FUN2MODEL ERC project under the supervision of Marta Kwiatkowska.
I hold a Ph.D. in Computer Science from Instituto Superior Técnico (IST), Universidade de Lisboa (UL), completed as part of a European Doctorate Programme in collaboration with the Czech Technical University (CTU) in Prague.
My doctoral research was supervised by Vasco Manquinho and Mikoláš Janota, with research hosted at INESC-ID, IST (Lisbon), and CIIRC, CTU (Prague).
During the final year of my M.Sc. in Computer Science, I was a Research Assistant at Carnegie Mellon University (CMU), collaborating with Ruben Martins. I also gained some industry experience as a Research Intern at OutSystems, a leading Portuguese software company.
I am the proud recipient of several 🏆 awards and grants 🏆, including the Vencer o Adamastor 2025 Award, which recognises outstanding young scientists in Portugal; the ACM SIGSOFT Distinguished Paper Award at FSE 2021; the 2025 Best PhD Thesis in AI in Portugal award; and the INESC-ID Best PhD Student 2025 Award.
📧 You can reach me at pmorvalho (AT) gmail.com 📧
PhD in Computer Science
Instituto Superior Técnico, Universidade de Lisboa, in collaboration with the Czech Technical University (CTU) in Prague
MSc in Computer Science and Engineering
Instituto Superior Técnico, Universidade de Lisboa
BSc in Computer Science and Engineering
Instituto Superior Técnico, Universidade de Lisboa
I am thrilled to share that I have received the 2025 Best PhD Thesis in Artificial Intelligence Award from APPIA, recognising doctoral research of outstanding merit in AI carried out at a Portuguese higher-education institution. 🎉 🎉
Excited to share that two of our papers have been accepted at EPIA 2026, covering the robustness of LLMs for code understanding and neuro-symbolic feedback for Vision-Language Models using MaxSAT! 🎉
Excited to share that three of our papers have been accepted at FLoC 2026, covering automated feedback for Prolog education, data-driven mutation testing for Prolog, and LLM-assisted MaxSAT modelling! 🎉
My research explores Neuro-symbolic AI, i.e., the synergy between Automated Reasoning (AR) and Machine Learning (ML), focusing on enhancing the robustness and reliability of ML models (e.g., Large Language Models), across diverse reasoning tasks, including but not limited to code understanding.
By integrating the mathematical rigor and precision of formal methods with the scalability and adaptability of machine learning, I strive to develop AI systems that are both reliable and efficient, tackling critical challenges at the frontier of trustworthy AI.
I am always excited to explore new ideas together! 📧 Feel free to get in touch 📧 if you are interested in collaborating! 😃
In this paper we assess whether SOTA LLMs can reason about Python programs or are simply guessing. We apply five semantics-preserving code mutations, which maintain program …
In this paper, we propose a neuro-symbolic approach that integrates formal constraint reasoning into the VLM solving process via a Maximum Satisfiability (MaxSAT) oracle. Rather …
In this paper, we study a neuro-symbolic approach in which an LLM translates a natural language description of an optimisation problem into executable Python code using PySAT. The …
In this talk, we will present an empirical study investigating whether state-of-the-art Large Language Models (LLMs) can genuinely reason about Python programs or are simply guessing. Using semantics-preserving code mutations and human expert analysis, we evaluate the reasoning quality and robustness of nine open-source and closed-access LLMs. Our results reveal substantial fragility to syntactic changes and show that even correct predictions can be based on flawed reasoning, raising important questions about the reliability of LLMs for programming tasks.
In this talk, we will present a neuro-symbolic approach that combines Vision-Language Models (VLMs) with MaxSAT reasoning to improve reliability on structured visual reasoning tasks. Using Sudoku as a case study, we show how a MaxSAT oracle can detect inconsistencies in VLM predictions and provide targeted feedback to guide the model toward correct solutions. Our results demonstrate the potential of combining neural perception with symbolic reasoning to build more reliable reasoning systems.
In this talk, we will present ProHelp, an automated assessment platform for Prolog built on top of GitSEED, and evaluate it in a large undergraduate logic programming course with 144 student responses. Our results show that students valued the feedback, with automatic testing standing out as the most useful type.